Professional Documents
Culture Documents
www.elsevier.com/locate/pla
Abstract
In experiments for single and coupled pendula, we demonstrate the effectiveness of a new control method based on dynamical systems theory
for stabilizing unstable aperiodic trajectories defined on infinite- or finite-time intervals. The basic idea of the method is similar to that of the OGY
method, which is a well-known, chaos control method. Extended concepts of the stable and unstable manifolds of hyperbolic trajectories are used
here.
© 2007 Elsevier B.V. All rights reserved.
Keywords: Control; Hyperbolic trajectory; Experiment; Stable and unstable manifolds; Infinite- and finite-time intervals
robotic manipulators and spacecrafts. Thus, our result is im- finite-time interval [t− , t+ ] in (2.1) is finite-time hyperbolic if
portant in engineering applications as well as in physical ones. Dx φ(γ̄ (t− ), t+ ; t− ) has no eigenvalue of a unit modulus. We
See [23] for the necessary technical details. define the Poincaré-type map for the finite-time hyperbolic tra-
jectory γ̄ (t) as
2. Hyperbolic trajectories
ψ(x) = φ x + γ̄ (t− ), t+ ; t− − γ̄ (t+ ). (2.3)
We first outline necessary information on hyperbolic trajec- If x(t) is a trajectory of (2.1), then the Poincaré-type map ψ(x)
tories on infinite- and finite-time intervals [23]. Consider gen- takes x(t− ) − γ̄ (t− ) to x(t+ ) − γ̄ (t+ ). Obviously, the origin
eral, non-autonomous vector fields of the form x = 0 is a hyperbolic fixed point of ψ , so that by a standard
ẋ = f (x, t; μ), x ∈ Rn , t ∈ R, μ ∈ Rm , (2.1) result of dynamical systems theory [1,2] it has stable and un-
stable manifolds, W s (0) and W u (0), on which orbits approach
where f is sufficiently smooth and μ is a parameter vector with (respectively depart from) the origin when ψ is iterated. We ex-
m, n > 0 some integers. tend the phase space Ē = {(x, t) ∈ Rn × [t− , t+ ]} and denote
Γ¯ = {(γ̄ (t), t) | t ∈ [t− , t+ ]}. We define the finite-time stable
2.1. Infinite-time interval case and unstable manifolds of Γ¯ as
We begin with the infinite-time interval (−∞, ∞). A trajec- W̄ s (Γ¯ ) = φ(x, t; t+ ), t ∈ Ē
tory x = γ (t) of (2.1) on (−∞, ∞) is said to be hyperbolic if
| x − γ̄ (t+ ) ∈ W s (0), t ∈ [t− , t+ ] ,
the associated linearized system
W̄ u (Γ¯ ) = φ(x, t; t− ), t ∈ Ē
ξ̇ = Dx f γ (t), t ξ, (2.2)
| x − γ̄ (t− ) ∈ W u (0), t ∈ [t− , t+ ] . (2.4)
has stable and unstable subspaces E s,u (τ ) such that trajectories
of (2.2) starting on E s (τ ) (respectively on E u (τ )) at t = τ ap- For τ ∈ [t− , t+ ] we also denote W̄ s,u (Γ¯ , τ ) = W̄ s,u (Γ¯ ) ∩ Στ ,
proach zero as t → +∞ (respectively t → −∞). We extend the which we call the finite-time stable and unstable slices of Γ¯ at
phase space of (2.1) to E = {(x, t) ∈ Rn × R} and define a time t = τ . Note that in its neighborhood, trajectories on the finite-
slice as Στ = {(x, t) ∈ E | t = τ }. We represent the hyperbolic time stable (respectively unstable) manifold approach (respec-
trajectory γ (t) as Γ = {(γ (t), t) | t ∈ R} in E . The hyperbolic tively depart from) the finite-time hyperbolic trajectory quickly
trajectory Γ has stable and unstable manifolds, W s (Γ ) and in the finite-time interval [t− , t+ ] if the stability (instability) is
W u (Γ ), such that trajectories of (2.1) starting on W s (Γ ) (re- strong. We can also prove that at time far from t± the finite-
spectively on W u (Γ )) approach Γ as t → +∞ (respectively time stable and unstable slices are O(e−c|t± | )-close to ones de-
t → −∞). See Fig. 1. We denote W s,u (Γ, τ ) = W s,u (Γ ) ∩ Στ , fined by assuming the vector field outside [t− , t+ ] [23]. Again,
which we call the stable and unstable slices of Γ at t = τ . The the finite-time stable and unstable slices, W̄ s,u (Γ¯ , τ ), are ap-
stable and unstable slices, W s,u (Γ, τ ), are well approximated proximately computed from the linearized system (2.2) with
by the stable and unstable subspaces, E s,u (τ ), which can be γ (t) = γ̄ (t).
easily computed in (2.2).
3. Control strategy
2.2. Finite-time interval case
Let γ (t) be an unstable hyperbolic trajectory of (2.1) on the
We turn to the case of a finite-time interval [t− , t+ ], where infinite-time interval (−∞, ∞). Using an idea similar to the
t− < 0 < t+ . Let φ(x0 , t; t0 ) be the flow generated by (2.1), OGY method [8], we apply a small perturbation in the direction
i.e., x(t) = φ(x0 , t; t0 ) is a solution of (2.1) satisfying an ini- of its unstable manifold to stabilize the hyperbolic trajectory
tial condition x(t0 ) = x0 . We say that a trajectory γ̄ (t) on the γ (t). Note that γ (t) is not assumed to be periodic as in usual
chaos control methods like the OGY method. We set the di-
mension m of the parameter vector such that it is the same as
the dimension nu of the unstable slice W u (Γ, τ ). Let the stable
and unstable subspaces E s,u (τ ) for the linearized system (2.2)
be spanned by vectors ejs,u (τ ) ∈ Rn , j = 1, . . . , ns,u . Introduce
a sequence of times, {tk | k ∈ Z}, with tk < tk+1 for any k.
We consider a trajectory x(t) deviating by a small amount
xk ∈ Rn from the hyperbolic trajectory γ (t) at t = tk and vary
the parameter value of μ ∈ Rm by μk ∈ Rm to shift it onto
the stable slice W s (Γ, tk+1 ) at t = tk+1 . Denote φk (x, μ) =
φ(x, tk+1 ; tk ; μ), where the dependence of φ on the parameter
μ is written explicitly. We linearize the flow φ to obtain
φk γ (tk ) + xk , μ + μk − γ (tk+1 )
Fig. 1. Stable and unstable manifolds of a hyperbolic trajectory. ≈ Dx φk γ (tk ), μ xk + Dμ φk γ (tk ), μ μk . (3.1)
224 K. Yagasaki / Physics Letters A 368 (2007) 222–226
Fig. 4. Control of the homoclinic trajectory for the single pendulum: (a) The Fig. 5. Control of the finite-time hyperbolic trajectory for the single pendulum:
angular displacement θ ; (b) the parameter variation μ. In panel (a) the dashed (a) the angular displacement θ ; (b) the parameter variation μ. In panel (a) the
curve represents the angular displacement when no feedback control force was dashed curve represents the angular displacement when no feedback control
applied, i.e., μ ≡ 0. force was applied, i.e., μ ≡ 0.
−1.816 × 10−4 and v h (10) ≈ 1.816 × 10−4 ), where e0s and We set θ∗ = −π , T = 10 and T = 1, and applied the
e0u are vectors spanning the stable and unstable subspaces, control method to the finite-time hyperbolic trajectory. Fig. 5
E s (+∞) and E u (−∞), respectively. The linearized maps shows an experimental result. We integrated the linearized sys-
Dx φk (x, μ) and Dμ φk (x, μ), k = −10, . . . , 9, were estimated tem for (4.1) numerically to obtain the Jacobian matrix of the
by numerically integrating (4.2) and Poincaré-type map, and approximated the finite-time stable and
unstable slices, W̄ s,u (Γ¯ , τ ), by their tangent spaces. Taking the
ξ̇1 = ξ2 , ξ̇2 = − cos θ h (t) ξ1 − δ1 ξ2 + 1, (4.3) experimental accuracy into account, we can state that the con-
respectively, over the interval [tk , tk+1 ]. The deviation x = trol method succeeded in stabilizing the trajectory while a small
( θ, v) from the target trajectory (θ h (t), v h (t)) was mea- deviation exists near t = 0.
sured at each time t = tk , and Eq. (3.2) is used to compute the
control force μk = μk . 4.3. Coupled pendula
Fig. 4 shows an experimental result on the dimensionless
time interval [−10, 10]. We see that the hyperbolic trajectory The dimensionless equation of motion for the coupled pen-
was stabilized, while the pendulum motion was very different dula is approximately given by
from it and converged to the stable equilibrium state (θ, v) =
θ̇1 = v1 ,
(0, 0) when the feedback control force was not applied. The
parameter variation is more or less large, compared with numer- v̇1 = − sin θ1 − δ1 v1 − δ0 sgn v1 − α(v1 − v2 ) + u1 (t),
ical simulation results in [23]. This is considered to be mainly θ̇2 = v2 ,
due to modeling errors for the experimental system. Thus, the
v̇2 = − sin θ2 − δ1 v2 − δ0 sgn v2 − α(v2 − v1 ) + u2 (t), (4.5)
control method is very robust.
We next consider a finite-time interval [−T , T ] and choose where α is a constant, and u1 (t) = u10 (t) + μ, u2 (t) = u20 (t),
(θ, v) = (θ0 (t; θ∗ ), v0 (t)) as our target trajectory, where for are control inputs such that (θ1 , v1 , θ2 , v2 ) = (θ0 (t; −π), v0 (t),
some T > 0 and θ∗ ∈ [−π, π) θ0 (t; 0), v0 (t)) is a solution of (4.5) when μ = 0. We can easily
⎧ show that the trajectory is finite-time hyperbolic, with nu = 1
)4 −4 T (t+T )3 ]
⎪
⎪ − π[(t+T 16T + θ∗ and choose it as our target. Fig. 6 shows an experimental result
⎪
⎪ 1 T
3
⎪
⎪ for control of the finite-time hyperbolic trajectory when α =
⎨ for t ∈ [−T , −T2 );
⎪
1.2, T = 10 and T = 1. The finite-time stable and unstable
θ0 (t; θ∗ ) = Tπ1 t + θ∗ + π for t ∈ [−T2 , T2 ]; (4.4)
⎪
⎪ slices, W̄ s,u (Γ¯ , τ ), were approximately estimated as in the case
⎪
⎪ π[(t−T ) +4 T (t−T ) ] + θ + 2π
4 3
⎪
⎪ ∗ of the single pendulum. Again, the control method succeeded
⎪
⎩ 16T1 T 3
in stabilization of the target.
for t ∈ (T2 , T ],
and v0 (t) = θ̇0 (t; θ∗ ), which is independent of θ∗ , with T1 = 5. Concluding remarks
T − T and T2 = T − 2 T . Here u(t) = u0 (t) + μ is de-
termined such that the target trajectory, which represents one Our experimental results show that the new control method,
rotation of the pendulum from θ = θ∗ at t = −T to θ = θ∗ + 2π which is considered to be an extension of the OGY method,
at t = T , is a solution of (4.1) when μ = 0. succeeded in stabilizing unstable trajectories on an infinite-
226 K. Yagasaki / Physics Letters A 368 (2007) 222–226
Acknowledgements
References